Research On Person Re-identification Based On Unsupervised Domain Adaptation

Person re-identification is a key research topic in the field of computer vision,and has a wide range of applications in real scenes,such as intelligent security,unmanned supermarkets,intelligent transportation,etc.Since supervised person re-identification relies on a large number of expensive labeled data,it is difficult to extend to real scenes.The unsupervised person reidentification technology has developed rapidly in recent years.As an important branch of the unsupervised domain,domain adaptation proposes to transfer the patterns emulated from the source domain to the unsupervised target domain,making the best of the rare labeled data to obtain a more generalized model.Among them,the cluster-based pseudo-label generation method has become the focus of scholars in the unsupervised adaptive person re-identification task because it can effectively alleviate the inter-domain differences.However,in the current domain adaptive person re-identification methods based on pseudo-label,there are many problems such as large inter-domain differences,poor quality of pseudo-label and weak feature representation capabilities.In response to the above issues,this paper major considers this topic from three aspects: rebuilding the mutual learning mode of dual networks,enriching and expanding the ability of feature representation,and enhancing the constraints of networks in feature space.The main research results of this article are as bellow:(1)This paper proposes an algorithm based on heterogeneous dual network(HDNet)mutual learning on the unsupervised domain adaptive.In order to solve the problem that two networks with the same structure are coupled with each other as the training progresses,this paper introduces the Transformer model.On the one hand,this model is used to improve the heterogeneity between the two networks,and to make up for the defect of the convolutional neural network in capturing global information on the other hand.Specifically,HDNet is composed of two heterogeneous networks,one of which uses convolution of limited receptive fields to obtain local information,and the other uses Transformer model to capture long-range dependency.Compared with the existing asymmetric structure,HDNet realizes the construction of long-range dependency between features.Experiments are carried out on three datasets widely used in person re-identification tasks,and the results prove that the proposed method is feasible.At the same time,a great quantity ablation experiments and visualization results further prove the improvement of this method on network coupling.(2)This paper proposes an algorithm based on local interaction and smooth alignment on the unsupervised adaptive.In order to extract more expressive features in the target domain,this paper proposes a part-pixel Transformer with smooth alignment fusion network(PTFNet).In view of the fact that most current Transformer-based methods directly divide the input image into a series of regularly spaced patches,which destroy the inherent attributes of the pedestrian in the image.This paper proposes a part-pixel Transformer(PPformer)module,which abandons the traditional segmentation method,retains more highly relevant regions,and models more fine-grained information.At the same time,pointing at the problem that the current multi-level feature fusion method ignores the potential relationship between features,this paper proposes a smooth alignment fusion(SAF)module to improve the representation ability of features by aggregating the smoothed low-level spatial information and high-level semantic information.Compared with the current mainstream methods,the effectiveness of PTFNet has been proved on three datasets widely used in person re-identification tasks,and the performance improvement of each module on the domain adaptive person re-identification model is testified by the ablation experiment.(3)This paper proposes a distribution consistency constraint with channel mutual-aware network(DCCMNet).In order to enhance the constraint between the probability distribution of the teacher model and the student model,this paper proposes a Distribution Consistency Loss(DCL)that does not rely on any pseudo-label.Its guiding model focuses more on the consistency of the probability distribution rather than the relationship between samples and labels.Moreover,for strengthening the semantic information of the model,an adaptive channel mutual-aware(ACMA)module is also proposed.This module consists of two branches,which are used to pay attention to the global and local information of the channel at the same time,and introduce the channel shuffle operation to realize the information perception between channels.Compared with the current mainstream methods,the effectiveness of DCCMNet has been proved on three datasets widely used in person re-identification tasks,and the desirability of each module is further proved by ablation experiment.To sum up,this paper mainly studies the domain adaptive tasks based on the pseudo-label method,and proposes three different domain adaptive networks: HDNet,PTFNet and DCCMNet,and experiments on multiple datasets have verified the advantages of this method.